Post-op chains silently reordered when the fused op is not first in the chain

[lwandrebeck]

Summary

The native loopers scan the post-op chain for the first fusable op —
relu / gelu_tanh / gelu_erf / sigmoid / tanh / swish — hand it to the
microkernel epilogue, and apply everything else afterwards via
apply_postops_tile(). The scan accepts a match at any index, but the epilogue
runs before the remaining ops, so fusing an op that is not at the head of the
chain silently reorders it.

binary_add:sigmoid computes binary_add(sigmoid(x)) instead of
sigmoid(binary_add(x)).

A binary/residual op followed by an activation is an ordinary pattern, and the
result is wrong numerics with nothing reported.

Not ISA-specific, and not tied to --no-aocldlp

Reproduced on an Intel Xeon 6767P with full AVX-512 (avx512f,
avx512_bf16, avx512_vnni), where the AVX-512 microkernels and epilogue are
the ones running — i.e. the default configuration on mainstream hardware.

Reproduce

Pin the chains through a gtest input file, with alpha=1 and beta=0 so this is
isolated from alpha/beta handling:

256,256,256,binary_add:sigmoid,native_gemm,false,false,1,0,,,tensor
256,256,256,sigmoid:binary_add,native_gemm,false,false,1,0,,,tensor
256,256,256,binary_mul:relu,native_gemm,false,false,1,0,,,tensor
256,256,256,relu:binary_mul,native_gemm,false,false,1,0,,,tensor
256,256,256,clip:tanh,native_gemm,false,false,1,0,,,tensor
256,256,256,tanh:clip,native_gemm,false,false,1,0,,,tensor

./gtests --op matmul --input_file F --ndims 2 --test 2 --seed 424242 \
         --gtest_filter="*TestMatmul.F32_F32/*"

chain	fusable op position	before	after fix
binary_add:sigmoid	second	FAIL	pass
binary_mul:relu	second	FAIL	pass
clip:tanh	second	FAIL	pass
sigmoid:binary_add	first	pass	pass
relu:binary_mul	first	pass	pass
tanh:clip	first	pass	pass

6 passed / 6 failed before, 12 passed / 0 failed after.

Suggested fix

Fuse only when the fusable op sits at index 0; otherwise leave it to the ordered
post-op pass. The same pattern appears in all four native loopers (fp32/bf16
GEMM, fp32/bf16 BRGEMM).

Reference branch: https://github.com/lwandrebeck/BullDNN/tree/upstream/postop-chain-reorder